281–290 of 1001 posts
I would bet all of my assets of my life that AGI will not be seen in the lifetime of anyone reading this message right now. That includes anyone reading this message long after the lives of those reading it on its post date have ended. Which of course raises the interesting question of how I can make good on this bet.
Should probably just short nvidia
>What takes the long amount of time and the way to think about it is that it’s a march of nines. Every single nine is a constant amount of work. Every single nine is the same amount of work. When you get a demo and something works 90% of the time, that’s just the first nine. Then you need the second nine, a third nine, a fourth nine, a fifth nine. While I was at Tesla for five years or so, we went through maybe three…
Earlier quoted context omitted.
back in the day alpha-beta search was AI hehe
As a young child in Indonesia we had an exceptionally fancy washing machine with all sorts of broken English superlatives on it, including "fuzzy logic artificial intelligence" and I used to watch it doing the turbo spin or whatever, wondering what it was thinking. My poor mom thought I was retarded.
>What takes the long amount of time and the way to think about it is that it’s a march of nines. Every single nine is a constant amount of work. Every single nine is the same amount of work. When you get a demo and something works 90% of the time, that’s just the first nine. Then you need the second nine, a third nine, a fourth nine, a fifth nine. While I was at Tesla for five years or so, we went through maybe three…
The interview which I've watched recently with Rich Sutton left me with the impression that AGI is not just a matter of adding more 9s. The interviewer had an idea that he took for granted: that to understand language you have to have a model of the world. LLMs seem to udnerstand language therefore they've trained a model of the world. Sutton rejected the premise immediately. He might be right in being skeptical here…
The fundamental idea behind temporal difference is that you can record any observable data stream over time and predict the difference between past and present based on your decision variables (e.g. camera movement, actuator movement, and so on). Think of it like the Minecraft clone called Oasis AI. The AI predicts the response to a user provided action.
Now imagine if it worked as presented. The data problem would be solved, because you are receiving a constant stream of data every single second. If anything, the RL algorithms are nowhere near where they need to be and continual learning has not been solved yet, but the best known way is through automatic continual learning ala Schmidhuber (co-inventor of LSTMs along with Hochreiter).
So, model based control is solved right? Everything that can be observed can be controlled once you have a model!
Wrong. Unfortunately. You still need the rest of reinforcement learning: an objective and a way to integrate the model. It turns out that reconstructing the observations is too computationally challenging and the standard computational tricks like U-Nets learn a latent representation that is optimized for reconstruction rather than for your RL objectives. There is a data exchange problem that can only realistically be solved by throwing an even bigger model at it, but here is why that won't work either:
Model predictive control tries to find the best trajectory over a receding horizon. It is inherently future oriented. This means that you need to optimize through your big model and that is expensive to do.
So you're going to have to take shortcuts by optimizing for a specific task. You reduce the dimension of the latent space and stop reconstructing the observations. The price? You are now learning a latent space for your particular task, which is less demanding. The dream of continual learning with infinite data shatters and you are brought down to earth: it's better than what came before, but not that much better.
Are "agents" just programs that call into an LLM and based on the response, it will do something?
Eg for a programming LLM with an agentic agent and access to a computer, would be able to, given design-doc.md and Todo.md, implement feature X, making sure it compiles, run some basic smoke tests, write appropriate unit tests, make sure they all pass, and finally push the code and create a draft PR.
Naturally, not every call into the agent is going to take the full 10 minutes. It may need to ask questions before getting started, or stop if there's an unrecoverable error. Sometimes you'll just need to tell it "continue", but the system should be capable of a 10-minute run (hopefully longer!) given enough support.
I would bet all of my assets of my life that AGI will not be seen in the lifetime of anyone reading this message right now. That includes anyone reading this message long after the lives of those reading it on its post date have ended. Which of course raises the interesting question of how I can make good on this bet.
Earlier quoted context omitted.
> yeah that "model of the world" would mean: babies are already born with "the model of the world" No, not necessarily. Babies don't interact with the world only by reading what people wrote wikipedia and stackoverflow, like these models are trained. Babies do things to the world and observe what happens. I imagine it's similar to the difference between a person sitting on a bicycle and trying to ride it, vs a person…
We’ve been thinking about reaching the singularity from one end, by making computers like humans, but too little thought has been given to approaching the problem from the other end: by making babies build their world model by reading Stack Overflow.
>What takes the long amount of time and the way to think about it is that it’s a march of nines. Every single nine is a constant amount of work. Every single nine is the same amount of work. When you get a demo and something works 90% of the time, that’s just the first nine. Then you need the second nine, a third nine, a fourth nine, a fifth nine. While I was at Tesla for five years or so, we went through maybe three…
I think a ton of people see a line going up and they think exponential. When in Reality, the vast majority of the time it’s actually logistic.
I always get a weird feeling when AI researchers and CS people start talking about comparisons between human brains and AI/computers Why is there a presumption that we (as people who have only studied CS) know enough about biology/neuroscience/evolution to make these comparisons/parallels/analogies? I enjoy the discussions but I always get the thought in the back of my head "...remember you're listening to 2 CS major…
>Why is there a presumption that we (as people who have only studied CS) know enough about biology/neuroscience/evolution to make these comparisons? Hubris.
Meanwhile, these systems translate languages, write code, play Go at superhuman levels, and pass medical licensing exams... all tasks you'd have sworn required "real understanding" a decade ago. At some point, look at the goddamn scoreboard. If you think there's something brains can do that these architectures fundamentally can't, name it specifically instead of gesturing vaguely at "inscrutability." The list of "things only biological brains can do" keeps shrinking, and your objection keeps sounding like "but my substrate is special!!1111"
I think it's a shame that a 146 minute podcast released ~55 minutes ago has so much discussion. Everybody here is clearly just reacting to the title with their own biases. I know it's against the guidelines to discuss the state of a thread, but I really wish we could have thoughtful conversations about the content of links instead of title reactions.
Be fair; plenty of people transcribe and read podcasts, and/or summarize/excerpt them.